05 — Running the app step-by-step

Duration: 12 min Prerequisites: chapter 02 (env + API key ready).

All commands

# From the project-2/ folder (where both repos sit side by side)
cd csv-llm-openai

# Quick start (helper script)
.\start.ps1

# OR manual
streamlit run app.py --server.port 8505

The app opens at http://127.0.0.1:8505.

The privacy banner

At the top of the page, a yellow warning:

⚠️ Privacy: every question sends a slice of your transactions to the OpenAI API. To stay 100% local, use csv-llm-ollama/ instead.

This is a key difference from Project 1. You’re immediately reminded data is leaving your machine.

First import

Same as Project 1:

1. Sidebar → CSV path = pre-filled with data/data1-anonymized.csv.
2. Profile name = data1.
3. Click 📥 (Re)load this CSV.
4. 734 transactions loaded into data1.

The Dashboards tab now shows the same 4 KPIs and 4 figures from Project 1.

Dashboards tab — same numbers as Ollama

KPI	Value
Total	$118,667.69
Monthly average	$9,888.97
Avg per transaction	$175.80
# expense transactions	675

Since the data and SQL queries are identical, the dashboards are byte-for-byte the same. Only the chat tab will behave differently.

Chat tab — the speed difference

Sidebar fields specific to this app:

• OPENAI_API_KEY (password mode). If empty, we read from environment or .env.
• Model: default gpt-4o-mini. You can switch to gpt-4o for max accuracy at ~10× the price.

Type Q2: What is my biggest spending category?

Typical answer in ~2 seconds:

Your biggest spending category is Housing with a total of $30,895.29.

⏱️ 1.8 s · 1 SQL call · ~$0.0003

And the 🔧 SQL #1 expander shows:

SELECT category, SUM(amount) AS total
FROM transactions
WHERE profile = 'data1' AND amount > 0
GROUP BY category
ORDER BY total DESC
LIMIT 1

category	total
Housing	30895.29

Identical SQL, identical answer. 5× faster than Project 1.

The 4 canonical questions

#	Question	Latency	Expected answer
Q1	How much did I spend on groceries last month?	1.5 s	Filter by category='Groceries' + last-month date
Q2	What is my biggest spending category?	1.8 s	Housing with $30,895.29
Q3	How many times did I visit Coastal Spoon Cafe this year?	2.0 s	19 times, total $1,221.34
Q4	List my top 5 merchants	2.5 s	Rent Payment Willow Street, StorageBox Monthly, GardenCare Subscription, Postbox Rental, Queenstown Path Pass

Total cost across all 4: ~$0.001 — one tenth of a cent.

Watch the live cost

Open platform.openai.com/usage in another tab and refresh after each question. You’ll see your spend increment by ~$0.0003 at a time. Reassuring.

Side-by-side comparison

If you also have Project 1 running on port 8504, you can compare answers in real time:

Aspect	Ollama (llama3.1:8b)	OpenAI (gpt-4o-mini)
Q2 latency (1 SQL call)	~5 s	~2 s
Q4 latency (1 SQL call)	~10 s	~3 s
SQL accuracy	Very good (95%)	Excellent (99%)
Answer style	Often verbose, explanatory	More concise
Cost for 4 questions	$0	~$0.001
Privacy	Data stays local	Data sent to OpenAI

Sequence diagram in the wild

If you click the debug toggle in the sidebar, each chat row prints the full sequence:

[14:32:01] User → "What is my biggest spending category?"
[14:32:01] LLM call #1 (gpt-4o-mini, 612 input tokens)
[14:32:02] tool_call: query_sql(SELECT category, SUM(amount) ...)
[14:32:02] safe_query → 1 row: ('Housing', 30895.29)
[14:32:02] LLM call #2 (gpt-4o-mini, 720 input tokens)
[14:32:03] Response: "Your biggest spending category is Housing..."
[14:32:03] Cost estimate: $0.00031

`gpt-4o` for max quality

For absolute precision (e.g. tricky multi-join questions, ambiguous time references), switch the sidebar model to gpt-4o. ~10× the cost, but rock-solid answers. Most of the time gpt-4o-mini is plenty.

Troubleshooting

Error	Cause	Fix
OPENAI_API_KEY missing	No .env or env var	Create csv-llm-openai/.env or paste the key in the sidebar
401 Unauthorized	Invalid or expired key	Regenerate at platform.openai.com
429 Too Many Requests	Quota / rate-limit	Wait 1 min, add credit, or upgrade plan
400 Bad Request: invalid tool call format	Old openai SDK	pip install --upgrade openai (≥ 1.40)
Empty Dashboards tab	Forgot to load the CSV	Click “(Re)load this CSV” in the sidebar
Port 8505 already in use	Previous Streamlit crashed	.\stop.ps1 then relaunch

Stop the app

.\csv-llm-openai\stop.ps1

Frees port 8505.

Takeaways

• Same UI, same dashboards, same numbers as Project 1.
• ~2 s per question vs ~10 s on Ollama (5× faster).
• Cost: under one cent for the entire course.
• Each question shows the SQL, the rows, and an estimated cost.

Next: Cost, latency, accuracy comparison →